Methods Inf Med 2003; 42(03): 287-296
DOI: 10.1055/s-0038-1634363
Original article
Schattauer GmbH

Entering the Black Box of Neural Networks

A Descriptive Study of Clinical Variables Predicting Community-Acquired Pneumonia
P. S. Heckerling
1   Department of Medicine, University of Illinois, Chicago, USA
,
B. S. Gerber
1   Department of Medicine, University of Illinois, Chicago, USA
2   Department of Bioengineering, University of Illinois, Chicago, USA
,
T. G. Tape
3   Department of Medicine, University of Nebraska, Omaha, Nebraska, USA
,
R. S. Wigton
3   Department of Medicine, University of Nebraska, Omaha, Nebraska, USA
› Author Affiliations
Further Information

Publication History

Received 07 June 2002

Accepted 10 October 2002

Publication Date:
07 February 2018 (online)

Summary

Objectives: Artificial neural networks have proved to be accurate predictive instruments in several medical domains, but have been criticized for failing to specify the information upon which their predictions are based. We used methods of relevance analysis and sensitivity analysis to determine the most important predictor variables for a validated neural network for community-acquired pneumonia.

Methods: We studied a feed-forward, back-propagation neural network trained to predict pneumonia among patients presenting to an emergency department with fever or respiratory complaints. We used the methods of full retraining, weight elimination, constant substitution, linear substitution, and data permutation to identify a consensus set of important demographic, symptom, sign, and comorbidity predictors that influenced network output for pneumonia. We compared predictors identified by these methods to those identified by a weight propagation analysis based on the matrices of the network, and by logistic regression.

Results: Predictors identified by these methods were clinically plausible, and were concordant with those identified by weight analysis, and by logistic regression using the same data. The methods were highly correlated in network error, and led to variable sets with errors below bootstrap 95% confidence intervals for networks with similar numbers of inputs. Scores for variable relevance tended to be higher with methods that precluded network retraining (weight elimination) or that permuted variable values (data permutation), compared with methods that permitted retraining (full retraining) or that approximated its effects (constant and linear substitution).

Conclusion: Methods of relevance analysis and sensitivity analysis are useful for identifying important predictor variables used by artificial neural networks.

 
  • References

  • 1 Metlay JP, Kapoor WN, Fine MJ. Does this patient have community-acquired pneumonia? Diagnosing pneumonia by history and physical examination. JAMA 1997; 278: 1440-5.
  • 2 Heckerling PS, Gerber BS, Tape TG, Wigton RS. Prediction of community-acquired pneumonia using artificial neural networks. Med Decis Making 2003; 23: 112-21.
  • 3 Penny W, Frost D. Neural networks in clinical medicine. Med Decis Making 1996; 16: 386-98.
  • 4 Tu JV. Advantages and disadvantages of using artificial neural networks versus logistic regression in predicting medical outcomes. J Clin Epidemiol 1996; 49: 1225-31.
  • 5 Faussett L. Fundamentals of Neural Networks: Architectures, Algorithms, and Applications. Englewood Cliffs, New Jersey: Prentice Hall; 1994
  • 6 van de Laar P, Heskes T, Gielen S. Partial retraining: a new approach to input relevance determination. Int J Neural Systems 1999; 9: 75-85.
  • 7 Hunter A, Kennedy L, Henry J, Ferguson I. Application of neural networks and sensitivity analysis to improved prediction of trauma survival. Comput Meth Prog Biomed 2000; 62: 11-9.
  • 8 Baxt WG. Analysis of the clinical variables driving decision in an artificial neural network trained to identify the presence of myocardial infarction. Ann Emerg Med 1992; 21: 1439-44.
  • 9 Heckerling PS, Tape TG, Wigton RS. et al. Clinical prediction rule for pulmonary infiltrates. Ann Intern Med 1990; 113: 664-70.
  • 10 Rumelhart DE, Hinton GE, Williams RJ. Learning internal representations by error propogation. In: Rumelhart DE, McClelland JL. eds. Parallel Distributed Processing: Explorations in the Microstructure of Cognition. Cambridge, MA: MIT Press; 1986: 318-64.
  • 11 McClelland JL, Rumelhart DE. Training hidden units. In: McClelland JL, Rumelhart DE. Eds. Explorations in Parallel Distributed Processing. Cambridge, MA: MIT Press; 1988: 121-60.
  • 12 Baxt WG, White H. Bootstrapping confidence intervals for clinical input variable effects in a network trained to identify the presence of acute myocardial infarction. Neural Computation 1995; 7: 624-38.
  • 13 Wolfram S. Mathematica: A system for doing mathematics by computer. Second Edition. Addison-Wesley Publishing Company, Reading, Massachusetts; 1991
  • 14 Freeman JA. Simulating Neural Networks with Mathematica. Reading, Massachusetts: Addison-Wesley Publishing Company; 1994
  • 15 Heckerling PS, Gerber BS. Feed-forward, back-propagation neural network modeling using Mathematica. Comput Meth Prog Biomed. (submitted for publication).
  • 16 Dorfman DD, Alf Jr E. Maximum likelihood estimation of parameters of signal detection theory and determination of confidence intervals - rating-method data. J Math Psych 1969; 6: 487-96.
  • 17 Heckerling PS. Parametric receiver operating characteristic (ROC) curve analysis using Mathematica. Comput Meth Prog Biomed 2002; 69: 65-73.
  • 18 Hinton GE. Connectionist learning procedures. In: Carbonell JG. ed. Machine Learning: Paradigms and Methods. Cambridge, MA: MIT Press; 1990: 185-234.
  • 19 Garson GD. Interpreting neural network connection weights. AI Expert 1991; 6: 47-51.
  • 20 Glorfeld LW. A methodology for simplication and interpretation of backpropagation-based neural network models. Expert Syst Appl 1996; 10: 37-54.
  • 21 Hosmer DW, Lemeshow S. Applied Logistic Regression. New York: John Wiley and Sons; 1989
  • 22 Engleman L. Stepwise logistic regression. In: Dixon WJ, Brown MB, Engleman L, Frane JW, Hill MA, Jennrich R. (ed). BMDP Statistical Software. Berkeley, California: University of California Press; 1985: 330-4.
  • 23 Efron B, Tibshirani RJ. An Introduction to the Bootstrap. New York: Chapman and Hall; 1993
  • 24 Baxt WG. Use of an artificial neural network for the diagnosis of myocardial infarction. Ann Intern Med 1991; 115: 843-8.
  • 25 Patil S, Henry JW, Rubenfire M, Stein PD. Neural network in the clinical diagnosis of acute pulmonary embolism. Chest 1993; 104: 1685-9.
  • 26 Tourassi GD, Floyd CE, Sostman HD, Coleman RE. Acute pulmonary embolism: artificial neural network approach for diagnosis. Radiology 1993; 189: 555-8.
  • 27 El-Solh AA, Hsiao CB, Goodnough S, Serghani J, Grant BJ. Predicting active pulmonary tuberculosis using an artificial neural network. Chest 1999; 116: 968-73.
  • 28 Lette J, Colletti BW, Cerino M. et al. Artificial intelligence versus logistic regression statistical modeling to predict cardiac complications after non-cardiac surgery. Clin Cardiol 1994; 17: 609-14.
  • 29 Diehr P, Wood RW, Bushyhead J, Krueger L, Wolcott B, Tompkins RK. Prediction of pneumonia in outpatients with acute cough: a statistical approach. J Chronic Dis 1984; 37: 215-25.
  • 30 Gennis P, Gallagher J, Falvo C, Baker S, Than W. Clinical criteria for the detection of pneumonia in adults: guidelines for ordering chest roentgenograms in the emergency department. J Emerg Med 1989; 7: 263-8.
  • 31 Singal BM, Hedges JR, Radack KL. Decision rules and clinical prediction of pneumonia: evaluation of low-yield criteria. Ann Emerg Med 1989; 18: 13-20.
  • 32 Begg CB, Greenes RA. Assessment of diagnostic tests when disease verification is subject to selection bias. Biometrics 1983; 39: 207-25.
  • 33 Gottschalk A, Hyzer C, Geer RT. A comparison of human and machine-based predictions of successful weaning from mechanical ventilation. Med Decis Making 2000; 20: 160-9.